# EXMOS: Explanatory Model Steering Through Multifaceted Explanations and Data Configurations

**Authors:** Aditya Bhattacharya, Simone Stumpf, Lucija Gosak, Gregor Stiglic, Katrien Verbert  
**arXiv:** [2402.00491](https://arxiv.org/abs/2402.00491)  
**Date:** 2024-02-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.00491v1/

## Key points

- Combining global data-centric and model-centric explanations is crucial for effective model steering in interactive machine-learning systems for healthcare experts.
- Incorporating local explanations such as counterfactual and what-if scenarios enhances the usefulness, understandability, and actionability of global explanations.
- Providing visual explanations with a layer of abstraction prevents information overload and facilitates user exploration.
- Detailed information on data issues should be included in secondary views to cater to decision-makers while maintaining simplicity in the main explanation dashboard.
- Disclosing the data collection process and allowing easy roll-back to previous configurations are essential features for promoting system adoption in high-stake domains like healthcare.
- Peer approval functionality can be beneficial in ensuring model steering through manual configurations by involving a panel of domain experts for review and validation.

## Summary

In our research on interactive machine-learning systems for healthcare experts, we found that combining global data-centric and model-centric explanations is crucial for effective model steering. By incorporating local explanations such as counterfactual and what-if scenarios, we can enhance the usefulness, understandability, and actionability of global explanations. It is also important to provide visual explanations with a layer of abstraction to prevent information overload and facilitate user exploration. Furthermore, detailed information on data issues should be included in secondary views to cater to decision-makers while maintaining simplicity in the main explanation dashboard. Disclosing the data collection process and allowing easy roll-back to previous configurations are essential features for promoting system adoption in high-stake domains like healthcare. Peer approval functionality can also be beneficial in ensuring model steering through manual configurations by involving a panel of domain experts for review and validation. Overall, our study highlights the significance of comprehensive explanations in guiding healthcare experts towards improving prediction models through automated and manual data configurations. By considering the design implications outlined in our research, interactive machine-learning systems can better support users in detecting and resolving potential data issues for model enhancement.